Last updated: July 2026. Most prompt engineering advice is folklore — magic words, threats, offers to tip. Modern models don’t need any of it. What still works is unglamorous: be specific, supply context, show an example, and iterate. Here’s the version that survives contact with actual work.
The four things that actually matter
1. Say what you want, precisely
Most bad output is a bad request. “Write about email marketing” has no constraints, so you get the average of everything ever written about email marketing. “Write 200 words explaining why open rates became unreliable after Apple’s Mail Privacy Protection, for a marketer who already knows the basics” has a length, an angle, an audience and a level. The second one can’t produce generic filler.
2. Give it the material
The single biggest quality jump comes from pasting in the actual context — the draft, the data, the transcript, the brand guidelines — rather than describing it. A model reasoning about your document beats a model reasoning about your summary of your document. This is also why context window size matters in practice rather than just on a spec sheet.
3. Show one example
If you want a particular format or voice, one example of it does more than three paragraphs describing it. “Match the tone of this” with a sample attached is the most reliable instruction in the whole discipline.
4. Treat it as a conversation
The first output is a draft, not a verdict. Say what’s wrong with it — “too formal”, “you’ve buried the point”, “cut the last paragraph” — rather than rewriting the prompt from scratch. Correcting is faster than re-specifying, and the thread keeps everything you’ve already established.
Techniques worth knowing
- Ask it to think first. On reasoning problems, asking for the working before the answer measurably improves the answer. Most models now have an explicit reasoning mode that does this for you.
- Assign a perspective, not a costume. “Review this as a security engineer would” changes what gets noticed. “You are a world-class genius” changes nothing.
- Specify the output shape. A table, JSON, five bullets under forty words each. Ambiguity about format is ambiguity you’ll fix by hand.
- Ask what’s missing. “What have I not considered here?” is the highest-yield question in this list and almost nobody asks it.
- Ask it to critique its own output. Then ask for the revision. It catches a surprising amount.
What to stop doing
- Politeness rituals and threats. No measurable effect. Offering a tip does nothing.
- Enormous prompt templates from social media. Long prompts dilute the instruction that mattered.
- Trusting confident output. Fluency is not accuracy — see below.
- Re-prompting from scratch when a correction would do.
The part that matters more than any technique
Every model will produce confident, well-formatted, wrong answers — including citations and statistics that don’t exist. No prompting technique fixes this. What fixes it is choosing a tool whose answers you can check: Perplexity cites its sources, and NotebookLM answers only from documents you upload, with citations back to the passage. Both are listings here rather than tested reviews, but they’re the right architecture for anything you’ll act on.
Where prompting stops and tooling starts
If you’re prompting the same thing repeatedly, stop prompting and build. Custom GPTs in ChatGPT and Projects in Claude hold your context and instructions so you don’t restate them. For writing specifically, a purpose-built tool with brand voice baked in beats a well-crafted prompt — see the writing tools comparison.
FAQs
Is prompt engineering still a skill worth learning?
The basics, yes — specificity and context. The elaborate technique collections have mostly been absorbed into the models themselves.
Do the same prompts work across different models?
The principles do. Exact wording that’s tuned to one model rarely transfers, which is a good reason not to over-invest in a specific incantation.
How long should a prompt be?
As long as the necessary context, and no longer. Length isn’t the goal; the material you supply is.
Get the best AI tools in your inbox
One useful email a week — tool verdicts, honest picks, and copy-paste prompts. No spam, unsubscribe anytime.